A Saudi laboratory has published a model of around 428 billion parameters aimed at Arabic, built from open weights of Chinese origin. The training did not start from scratch: an existing base was specialised. Three threads we follow separately converge here: underserved languages, the spread of open weights, and the geopolitics of compute.
We wrote on Wednesday about the languages that major models ignore that regional players were taking up the issue by starting from open bases. Here is a textbook case, three days later.
The mechanism, and why it changes everything
Training a model of several hundred billion parameters from scratch costs hundreds of millions of dollars and requires skills that very few teams possess. That is what reserved the production of cutting-edge models for a handful of players, as we explained in our article on foundation models.
Starting from an open base and specialising it costs far less. You keep the general capabilities acquired by the original laboratory, and you add what it lacked: a corpus in a specific language, cultural references, local usage.
That is the logic we described for Thomson Reuters' legal model: the foundation is public, the specialisation is proprietary. Applied to a language rather than a profession, it produces the same result — a player who could never have built a cutting-edge model gets one.
Why Arabic
Arabic is spoken by several hundred million people, across around twenty countries, with considerable regional variants. It is structurally poorly served by Western models, for the reasons we have explained: imbalanced corpus, tokenizer optimised for English, lack of cultural context.
And the Gulf has what many other regions lack: substantial capital and a firm political will to have national capabilities. This is a case where need, means, and intent coincide.
Look at the chain: a model designed in China, released under a permissive licence, taken up in Saudi Arabia, specialised for a language that neither American nor Chinese laboratories serve well. No Western player is involved. This is exactly the effect we anticipated in our article on countries told to pick a side: in a world of blocs, open weights are the only objects that cross borders without permission. And China's strategy of mass publication produces its most concrete effect here — influence where China is not.
The caveats
No independent evaluation. A model of this size on a language poorly covered by standard tests is particularly hard to measure. The announced performance comes from the laboratory itself.
Dependency shifts; it does not disappear. A model built on a Chinese base inherits what that base contains: its choices, its limits, its possible biases. We explained this regarding distillation: a student does not surpass its teacher, and it inherits its flaws. The sovereignty gained is partial.
The licensing question. Depending on the terms of the base used, obligations differ. This is the point we detailed in our article on licences, and it applies here with political stakes.
What to take away
This model illustrates a transformation that three months of news have made clear. The question is no longer who builds the best models, but who can adapt them. And that capability is far more widely distributed.
This produces an AI map far more fragmented than the two blocs usually described. Any region with capital and a linguistic need can now equip itself with a serious model by building on what others have published. It is not independence, but it is no longer dependence on a single supplier, and for many countries that is already a considerable change.